Lithium-ion battery state-of-health estimation for fragmented charging segments in real-world fleets

To estimate lithium-ion battery state of health (SOH) from only a single charging segment in real-world fleets, this paper proposes a two-stage SOH estimation framework named HG–HLAT. Real-world electric vehicle data usually consist of single charging segments with variable lengths, random start and end SOC values, and noise. These data are hard to use with traditional methods that rely on complete cycles or long historical trajectories. To address this issue, we first build a histogram-statistics-based Gramian angular field representation, named Histogram-GAF (HG). It mines the voltage, current, and temperature distributions and compresses irregular charging segments into multi-channel GAF heatmaps based on statistical distributions. We then design a Hybrid Latent Attention Transformer (HLAT). It combines convolutional pre-encoding, alternating low-rank latent and standard full-rank attention, and an attention-weighted pooling regression head to extract degradation-sensitive features efficiently. We evaluate HG–HLAT on batteries with different operating conditions and chemistries, including 139 fast-charging cells under laboratory conditions and operational data from 260 in-service EVs. HG–HLAT achieves the best RMSE of 1.17% on the laboratory test set and an RMSE of 2.00% on the real-world EV fleet test set against offline constructed pack-equivalent SOH proxy labels. It also shows low relative inference cost and modest memory demand on the tested GPU platform. These results indicate that the method can effectively track degradation-related information in fragmented charging segments and has application potential for fleet-scale, vehicle–cloud collaborative, or edge-assisted SOH estimation.